Since August 11, 2026, Anthropic has embedded an invisible AI watermark in text produced by Claude, and signs generated files (images, SVG) with provenance metadata. The move responds to an EU AI Act obligation that took effect on August 2, 2026. For a small business that uses Claude daily to write, translate, or illustrate, here is what actually changes.
In brief
- Since August 11, 2026, Anthropic has been embedding an invisible watermark in text generated by Claude models launched on or after August 2, 2026, worldwide.
- Generated files (.svg, .png, .jpg) receive signed metadata under the C2PA standard, showing whether Claude processed the file.
- The move implements the Code of Practice on Transparency of AI-Generated Content under the EU AI Act (Article 50), in force since August 2, 2026, signed by roughly 190 organizations including Google, Meta, Microsoft, Mistral AI, and OpenAI.
- Anthropic is upfront about the limits: a detected watermark is a signal, not proof of authorship; it can disappear after heavy editing, a format conversion, or a screenshot.
- Failing to comply with Article 50 can lead to a fine of up to 15 million euros or 3% of global annual turnover.
What Anthropic announced on August 11, 2026
Anthropic confirmed it is rolling out a marking system for all content generated by Claude. Two distinct mechanisms are in place, one for text, one for files.
For text, an invisible pattern is woven directly into the generated response. It does not change the meaning, quality, or readability of the content. Because it is part of the text itself, it survives copy-and-paste and may persist through some editing.
For files (SVG, PNG, JPG images produced by Claude), Anthropic attaches signed metadata following the open C2PA standard (Coalition for Content Provenance and Authenticity), already used by Adobe, Microsoft, and Google. This metadata shows that a file was processed by Claude and can reveal later tampering.
Marking applies to Claude models launched on or after August 2, 2026, and Anthropic says it is working to extend it retroactively to earlier models. Notably, although the rule comes from the European Union, Anthropic applies it everywhere in the world, not only to users located in Europe.
Why now: the EU AI Act's Transparency Code
This announcement follows directly from Article 50(2) of the EU AI Act, which requires providers of generative AI systems to mark their content in a machine-detectable way. This transparency provision took effect on August 2, 2026, alongside most of the regulation.
To help companies comply, the European Commission published a Code of Practice on Transparency of AI-Generated Content. Roughly 190 organizations signed it ahead of the deadline, including Google, Meta, Microsoft, Mistral AI, and OpenAI. Anthropic signed as both a model provider (Claude) and a provider of systems built on that model.
August 2, 2026
Article 50 takes effect
Late July 2026
190 signatories
August 11, 2026
Anthropic's announcement
Text watermark and file metadata: two different logics
The two mechanisms work differently and do not cover the same use cases. The table below summarizes the differences that matter for a small business.
| Criterion | Text watermark | File metadata (C2PA) |
|---|---|---|
| Content covered | Claude text responses | Generated images: SVG, PNG, JPG |
| Survives copy-paste | Yes | No (metadata lost outside the file) |
| Visible to the eye | No, invisible | No, embedded in file data |
| Survives format conversion | Partially, depending on editing | No, lost on reconversion |
| What it proves | Claude likely processed this text | Claude processed this file, no tampering flagged |
Key takeaway
A detected watermark does not prove that a text was written entirely by Claude. It can appear on human-written text that Claude only edited, translated, summarized, or reformatted. Conversely, heavy editing or very short text may leave no detectable signal.
What this means in practice for a small business
Most small businesses use Claude as deployers, not model providers: they use it to write emails, marketing content, reports, or visuals, not to build their own AI. Three practical consequences follow from Anthropic's announcement.
First, content that passed through Claude can now be identified as such, even after being copied into a CMS, an email tool, or a presentation. A business that publishes AI-written content without substantially reworking it needs to factor this in, especially for public-facing content (website, social media, press releases).
Second, Article 50 of the EU AI Act also places obligations on deployers, not only on model providers. A small business that publishes AI-generated text or images that resemble authentic content (a realistic photo, a news article, a synthetic voice) must generally disclose that it was AI-generated or manipulated, with limited exceptions for clearly creative or artistic use, or when there has been substantial human editorial review.
Third, this announcement strengthens traceability in case of disputes: a company that can prove a visual came from Claude, via an intact C2PA metadata trail, has a technical piece of evidence when the origin of content is challenged.
Before the announcement
No simple way to know whether text came from Claude once copied outside the interface. Generated images carried no verifiable trace after export.
Since August 11, 2026
Text keeps an invisible watermark after copy-paste. Images keep a signed C2PA trail as long as they are not reconverted or captured as a screenshot.
A simple checklist for small businesses
No technical skills are needed to adapt to this change. A few habits are enough: review your communication policy on AI use, substantially edit and enrich content meant for the public rather than publishing raw Claude output, and inform marketing or communications teams that this marking now exists and can be detected by third-party tools, including by curious clients or journalists.
This move fits a broader trend: transparency about the origin of AI content is becoming a de facto standard, driven by European regulation but applied worldwide by major providers. With measured optimism, this is good news for customer trust toward companies that use AI responsibly and openly.
FAQ
Is Claude's watermark visible in the text?
No. It is an invisible pattern embedded in the text's structure, which does not change the meaning or readability of the response. It can only be detected by specific tools, not by eye.
Does a small business need to disclose that it uses Claude for its content?
It depends on the context. Article 50 of the EU AI Act requires disclosure for content that mimics reality (a realistic image, voice, or news-like content) that is generated or heavily modified by AI, unless it has been substantially reworked by a human or is clearly artistic.
Does the watermark work on every Claude model?
It applies by default to models launched on or after August 2, 2026. Anthropic said it is working to extend it retroactively to earlier models, without a precise timeline announced so far.
What happens if Claude's content is heavily edited afterward?
The text watermark can persist through light editing, but a major rewrite, a translation, or very short text may leave no detectable signal. For files, reconverting the format or taking a screenshot erases the C2PA metadata.
To learn more about EU AI Act obligations for small businesses, see our guide AI Act: what changes on August 2, 2026 for SMEs and browse all our resources on AI in business.
